LLaDA2.1: Speeding Up Text Diffusion via Token Editing
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915796822786048 |
|---|---|
| author | Bie, Tiwei Cao, Maosong Cao, Xiang Chen, Bingsen Chen, Fuyuan Chen, Kun Du, Lun Feng, Daozhuo Feng, Haibo Gong, Mingliang Gong, Zhuocheng Gu, Yanmei Guan, Jian Guan, Kaiyuan He, Hongliang Huang, Zenan Jiang, Juyong Jiang, Zhonghui Lan, Zhenzhong Li, Chengxi Li, Jianguo Li, Zehuan Liu, Huabin Liu, Lin Lu, Guoshan Lu, Yuan Ma, Yuxin Mou, Xingyu Pan, Zhenxuan Qiu, Kaida Ren, Yuji Tan, Jianfeng Tian, Yiding Wang, Zian Wei, Lanning Wu, Tao Xing, Yipeng Ye, Wentao Zha, Liangyu Zhang, Tianze Zhang, Xiaolu Zhao, Junbo Zheng, Da Zhong, Hao Zhong, Wanli Zhou, Jun Zhou, Junlin Zhu, Liwang Zhu, Muzhi Zhuang, Yihong |
| author_facet | Bie, Tiwei Cao, Maosong Cao, Xiang Chen, Bingsen Chen, Fuyuan Chen, Kun Du, Lun Feng, Daozhuo Feng, Haibo Gong, Mingliang Gong, Zhuocheng Gu, Yanmei Guan, Jian Guan, Kaiyuan He, Hongliang Huang, Zenan Jiang, Juyong Jiang, Zhonghui Lan, Zhenzhong Li, Chengxi Li, Jianguo Li, Zehuan Liu, Huabin Liu, Lin Lu, Guoshan Lu, Yuan Ma, Yuxin Mou, Xingyu Pan, Zhenxuan Qiu, Kaida Ren, Yuji Tan, Jianfeng Tian, Yiding Wang, Zian Wei, Lanning Wu, Tao Xing, Yipeng Ye, Wentao Zha, Liangyu Zhang, Tianze Zhang, Xiaolu Zhao, Junbo Zheng, Da Zhong, Hao Zhong, Wanli Zhou, Jun Zhou, Junlin Zhu, Liwang Zhu, Muzhi Zhuang, Yihong |
| contents | While LLaDA2.0 showcased the scaling potential of 100B-level block-diffusion models and their inherent parallelization, the delicate equilibrium between decoding speed and generation quality has remained an elusive frontier. Today, we unveil LLaDA2.1, a paradigm shift designed to transcend this trade-off. By seamlessly weaving Token-to-Token (T2T) editing into the conventional Mask-to-Token (M2T) scheme, we introduce a joint, configurable threshold-decoding scheme. This structural innovation gives rise to two distinct personas: the Speedy Mode (S Mode), which audaciously lowers the M2T threshold to bypass traditional constraints while relying on T2T to refine the output; and the Quality Mode (Q Mode), which leans into conservative thresholds to secure superior benchmark performances with manageable efficiency degrade. Furthering this evolution, underpinned by an expansive context window, we implement the first large-scale Reinforcement Learning (RL) framework specifically tailored for dLLMs, anchored by specialized techniques for stable gradient estimation. This alignment not only sharpens reasoning precision but also elevates instruction-following fidelity, bridging the chasm between diffusion dynamics and complex human intent. We culminate this work by releasing LLaDA2.1-Mini (16B) and LLaDA2.1-Flash (100B). Across 33 rigorous benchmarks, LLaDA2.1 delivers strong task performance and lightning-fast decoding speed. Despite its 100B volume, on coding tasks it attains an astounding 892 TPS on HumanEval+, 801 TPS on BigCodeBench, and 663 TPS on LiveCodeBench. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_08676 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | LLaDA2.1: Speeding Up Text Diffusion via Token Editing Bie, Tiwei Cao, Maosong Cao, Xiang Chen, Bingsen Chen, Fuyuan Chen, Kun Du, Lun Feng, Daozhuo Feng, Haibo Gong, Mingliang Gong, Zhuocheng Gu, Yanmei Guan, Jian Guan, Kaiyuan He, Hongliang Huang, Zenan Jiang, Juyong Jiang, Zhonghui Lan, Zhenzhong Li, Chengxi Li, Jianguo Li, Zehuan Liu, Huabin Liu, Lin Lu, Guoshan Lu, Yuan Ma, Yuxin Mou, Xingyu Pan, Zhenxuan Qiu, Kaida Ren, Yuji Tan, Jianfeng Tian, Yiding Wang, Zian Wei, Lanning Wu, Tao Xing, Yipeng Ye, Wentao Zha, Liangyu Zhang, Tianze Zhang, Xiaolu Zhao, Junbo Zheng, Da Zhong, Hao Zhong, Wanli Zhou, Jun Zhou, Junlin Zhu, Liwang Zhu, Muzhi Zhuang, Yihong Machine Learning Artificial Intelligence While LLaDA2.0 showcased the scaling potential of 100B-level block-diffusion models and their inherent parallelization, the delicate equilibrium between decoding speed and generation quality has remained an elusive frontier. Today, we unveil LLaDA2.1, a paradigm shift designed to transcend this trade-off. By seamlessly weaving Token-to-Token (T2T) editing into the conventional Mask-to-Token (M2T) scheme, we introduce a joint, configurable threshold-decoding scheme. This structural innovation gives rise to two distinct personas: the Speedy Mode (S Mode), which audaciously lowers the M2T threshold to bypass traditional constraints while relying on T2T to refine the output; and the Quality Mode (Q Mode), which leans into conservative thresholds to secure superior benchmark performances with manageable efficiency degrade. Furthering this evolution, underpinned by an expansive context window, we implement the first large-scale Reinforcement Learning (RL) framework specifically tailored for dLLMs, anchored by specialized techniques for stable gradient estimation. This alignment not only sharpens reasoning precision but also elevates instruction-following fidelity, bridging the chasm between diffusion dynamics and complex human intent. We culminate this work by releasing LLaDA2.1-Mini (16B) and LLaDA2.1-Flash (100B). Across 33 rigorous benchmarks, LLaDA2.1 delivers strong task performance and lightning-fast decoding speed. Despite its 100B volume, on coding tasks it attains an astounding 892 TPS on HumanEval+, 801 TPS on BigCodeBench, and 663 TPS on LiveCodeBench. |
| title | LLaDA2.1: Speeding Up Text Diffusion via Token Editing |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2602.08676 |